Cross-user variability remains a fundamental challenge in wearable Human Activity Recognition (HAR), as differences in body morphology and movement styles cause significant distribution shifts across users. Existing approaches typically focus on feature-level alignment while overlooking the biomechanical invariants-shared anatomical constraints governing human movement that persist across individuals. Moreover, current Graph Neural Network (GNN)-based HAR methods rely on single-type edge relationships (e.g., physical adjacency) with static edge features, limiting their capacity to capture the diverse and dynamic biomechanical dependencies essential for cross-user generalization. We propose Edge-Enhanced Graph-Based Adversarial Domain Generalization (EEG-ADG), a novel framework that explicitly encodes multi-relational anatomical knowledge into graph-based representation learning with probabilistic edge modeling. Unlike prior work that treats sensor relationships as fixed and homogeneous, EEG-ADG constructs a unified graph capturing three complementary biomechanical priors: Interconnected Units, Analogous Units, and Lateral Units. To handle the inherent uncertainty in these relationships across users, we introduce a Variational Edge Feature Extractor that learns probabilistic representations of inter-sensor correlations-a capability absent in deterministic edge-based GNNs. An adaptive edge attention mechanism further modulates relationship importance based on activity context, while adversarial training on graph-structured representations disentangles activity-relevant features from user-specific biases. Comprehensive experiments on the OPPORTUNITY and DSADS benchmarks demonstrate that EEG-ADG achieves state-of-the-art cross-user generalization performance, outperforming existing domain generalization methods. Our findings highlight the value of integrating biomechanical prior knowledge with probabilistic graph learning for robust, generalizable HAR systems.
In human-centric manufacturing, collaborative robots (cobots) are increasingly expected to continually acquire knowledge as products and workflows evolve, and learning from human demonstration videos has emerged as a flexible means to this end. Realizing this paradigm requires perception that can be deployed with minimal hardware overhead and that generates executable task knowledge for robotic manipulation. Existing video-based methods rarely satisfy both requirements. Geometric pipelines rely on CAD models or multi-camera rigs, hindering scalable deployment across diverse workpieces, while pre-trained vision-language models require demonstration corpora rarely available in high-mix, low-volume manufacturing. In addition, existing methods recover only low-level geometry or high-level semantic labels in isolation, leaving a gap between geometric perception and executable task knowledge. To address these limitations, we propose Video2Knowledge, a video-based learning framework that takes a single monocular RGB video as input and distills it into temporally consistent task knowledge, covering action sequences, object state transitions, and spatial relations, without requiring additional sensors, manual CAD models, or task-specific retraining. The framework couples a physics-aware 6D object pose corrector with hand-object interaction analysis, and consolidates the results into a structured knowledge representation that maps directly to robot action sequences. Experimental evaluation on the HA-ViD and Assembly101 datasets demonstrates that the physics-aware correction yields improved pose coherence, while the extracted task knowledge faithfully recovers demonstrated action semantics and relative geometric configurations. A real-world case study further validates the transfer of task knowledge from a monocular demonstration video to physical cobot execution, confirming the practical deployability in manufacturing scenarios.
Geoprivacy is often framed as a problem of protecting coordinates, masking traces, or managing user consent. This framing is now insufficient. In an era of pervasive datafication, AI-enabled inference, platform dependence, and corporate control over location and mobility data, the central question is no longer whether exact locations can be hidden, but whether meaningful control over location-derived knowledge remains possible at all. This paper asks: have we already lost geoprivacy? We argue that geoprivacy has not disappeared, but its conditions have fundamentally changed. Location data are now continuously produced through everyday digital participation, linked across platforms and modalities, and transformed by machine learning, large language models, and agentic systems into sensitive inferences about identity, behavior, association, vulnerability, and future action. At the same time, risks and protections are unevenly distributed: privacy increasingly functions as socioeconomic, political, and institutional privilege, while marginalized communities, sensitive places, non-human environments, and collective forms of knowledge face heightened exposure and risk. These challenges are intensified by corporate gatekeeping, regulatory lag, harmful defaults, limited transparency, and epistemic blind spots within existing geoprivacy norms and directions. We therefore reposition geoprivacy as a socio-technical, legal, institutional, and political problem, not merely a technical one. Looking forward, we outline responsibilities for policymakers, researchers, communities, and industry, including privacy-protective defaults, regulation of spatial inference, public-good data governance, geoprivacy literacy, cyber-threat exercises, and community-controlled infrastructures. Rather than declaring geoprivacy lost, we argue that it must be continually renegotiated as a moving boundary, one that requires GI science to act not only as a technical field, but as a bridge between spatial reasoning, governance, justice, and democratic control over location-based data and futures.
The global drive toward clean energy has increased interest in green hydrogen as a carbon-free fuel. Proton exchange membrane water electrolysis is efficient and renewable-compatible, but designing cost-effective hydrogen evolution reaction (HER) and oxygen evolution reaction (OER) electrocatalysts remains challenging, with most machine learning models lacking transparent, generalizable selection rules. In this study, we present a Rough Set Machine Learning (RSML) framework to predict Gibbs free energy of adsorption (Delta G) for bimetallic alloy catalysts based on decision attributes from a high-throughput density functional theory (DFT) dataset. The data set includes 21 physicochemical, structural, and catalytic descriptors, spanning fundamental periodic properties, electronic structure parameters, and reaction thermodynamics to predict and classify highperformance bimetallic alloys for HER and OER. The model generates interpretable IF-THEN rules to classify conditional attributes and identify optimal catalyst configurations. RSML feature selection identified four core attributes for HER (site nature, site atom, crystalline structure, facets) and three for OER (site nature, site atom, crystalline structure). Rules with relative strength greater than 35 were retained, resulting in 64 for HER and 69 for OER. The trained RSML achieved 85.43% accuracy for HER, outperforming all tested black-box models, and 91.18% for OER, performing competitively with leading ensemble models. The approach identified several previously underexplored bimetallic combinations with balanced adsorption-desorption energetics adhering to the Sabatier principle, suggesting their potential as candidates for further experimental evaluation toward green hydrogen generation. The findings show RSML's capability to provide transparent, scientifically coherent rules, enabling rational, high-accuracy catalyst design for green hydrogen production.
Efficient photocatalytic CO2 reduction under low-concentration conditions remains a challenge due to severe mass transfer limitations and sluggish activation kinetics. Herein, a defect-enabled S-scheme capture-conversion catalytic architecture is rationally constructed by coupling S-vacancy-engineered ZnIn2S4 with TiO2 (Vs-ZIST) for selective photocatalytic CO2 reduction to CH4 under a sub-ambient absolute pressure of 40 kPa. In this synergistic framework, TiO2 acts as an interfacial capture center to enrich CO2 molecules, while the sulfur-vacancyrich ZnIn2S4 serves as the electron-rich domain driving CO2 reduction. Consequently, the optimized Vs-ZIST photocatalyst delivers a CH4 evolution rate of 104.8 mu mol g(-1) with selectivity of 80.2%. Comprehensive in situ spectroscopic investigations and theoretical calculations elucidate that the S-scheme pathway facilitates the spatial separation of photogenerated charge carriers while preserving redox potentials, steering electron accumulation onto the vacancy sites. Concurrently, sulfur vacancies modulate the local electronic structure, stabilizing key *CO intermediates to thermodynamically favor subsequent hydrogenation toward CH4. This study elucidates defect-assisted S-scheme charge transfer and provides design guidelines for efficient photocatalytic CO2 conversion under sub-ambient absolute CO2 pressures.
